A gossip protocol for subgroup multicast

Abstract

Gossip-based multicast can be an effective tool for providing highly reliable and scalable message dissemination. We consider the problem of gossiping within overlapping process groups. If each subgroup independently runs a uniform gossip protocol, then the total gossip overhead could be high for a process that is a member of many subgroups. We present a novel gossip protocol that allows individual subgroup members to trade-off update quality for gossip overhead, enabling processes to belong to several subgroups while maintaining a low total gossip overhead. Our results include a mathematical model for message dissemination under this modified gossip protocol, and an algorithm that computes gossip parameters such that all processes within a subgroup achieve their desired update quality.